UniReg | Medical Image Registration

One model. Diverse registration scenarios.

Moving beyond independently trained networks for specific anatomical regions, modalities, and registration types.

UniReg is a conditional unified framework that estimates deformation fields across heterogeneous CT and MR registration tasks, by jointly modeling anatomical priors, inter-/intra-subject constraints, and instance-specific image features.

Original motivation figure from the UniReg paper — Motivation figure reproduced from the final manuscript.

Method Overview

Condition-aware registration with a unified model architecture.

UniReg complete framework architecture diagram — UniReg framework overview (original figure from the manuscript).

01 / STRUCTURE

Anatomical structure priors

Provides anatomical guidance for alignment across different organs and imaging settings.

02 / TASK

Registration type constraints

Conditions the deformation estimation process on inter-subject and intra-subject registration requirements.

03 / INSTANCE

Instance-specific features

Adapts the predicted deformation field to the characteristics of each moving and fixed image pair.

Experimental Results

Comprehensive comparisons across heterogeneous CT and MR registration scenarios.

CT + MRMultiple modalities

6 tasksJoint training evaluation

3 variantsCNN · C2F · MLP

Quantitative registration accuracy

Selected Dice similarity coefficient (DSC) results from the final manuscript's six-task comparison, reported as percentages. The strongest method varies by task.

Method	HeadNeck	Chest	Abdomen	Liver	Cardiac MR	Brain MR
UniReg (C2F)	57.04	56.69	55.08	87.08	76.10	81.79
UniReg (MLP)	57.02	56.48	55.32	87.03	76.21	81.97
DEEDs	54.21	52.72	46.52	83.50	75.01	73.84

Values are DSC (%) transcribed from the paper's principal comparison table. Consult the paper for complete baseline comparisons, deformation regularity metrics, and evaluation details.

Ablation Study: Single-task vs. Joint Training

Table IV compares training on each registration task independently with joint training using the same backbone and training protocol. Joint training improves Dice similarity coefficient (DSC) across all six tasks.

Dataset	Single-task DSC (%)	Joint-training DSC (%)	Improvement
HeadNeck	53.00	57.04	+4.04
Chest	51.62	56.69	+5.07
Abdomen	52.01	55.08	+3.07
Liver	84.79	87.08	+2.29
Cardiac MR	75.37	76.10	+0.73
Brain MR	81.77	81.79	+0.02

Table IV. Comparison between single-task training and joint training on different datasets. Both settings use the same backbone and training protocol. Values are DSC (%), and improvements are measured in percentage points.

Qualitative registration examples

Qualitative registered CT slices and segmentation overlays — Representative alignment results and segmentation overlays, including highlighted registration artifacts (original paper figure).

Efficiency vs. registration accuracy

Training efficiency and registration accuracy comparison — Training efficiency–accuracy comparison (original paper figure).

Citation

If UniReg is useful for your research, please consider citing our work.

@article{li2026unireg,
  title={UniReg: Conditional Unified Model for Medical Image Registration},
  author={Li, Zi and Zhang, Jianpeng and Ma, Tai and Mok, Tony C. W. and Zhou, Yan-Jie and Chen, Zeli and Ye, Xianghua and Lu, Le and Chen, Cheng and Jin, Dakai},
  journal={IEEE Transactions on Circuits and Systems for Video Technology},
  year={2026},
  note={Accepted, publication metadata pending}
}

Bibliographic volume, pages and DOI will be updated when the final publication record is available.